Building libsamplerate with icc 12 (performance problem)
I'm using "libsamplerate" library. When I rebuild it with every new Intel compiler's version the lib performance is degrading like 10%. So now with icc 12 the lib works about 30% slower comparing to old icc 9 build. What am I doing wromg? Are there some compiler's option I should check?
The question is probably more appropriate to the mail list of that application, but pointless if you aren't willing to say what options you are using. A majority of the vectorization performance regressions I have seen in icl 12.0 can be dealt with by using CEAN syntax; needless to say, that's not a satisfactory solution. In such cases, you would have seen changes in opt-report (or vec-report, which is included in opt-report). The default for icl 32-bit compilation changed from /arch:IA32 (no vectorization) to /arch:SSE2 (including auto-vectorization) in icl 10.0. If you mix float and double, IA32 may perform better than SSE2.